Sakana Fugu: The Multi-Agent AI System Beating Frontier Models
How a Tokyo-based system routes tasks across GPT, Claude, and Gemini to outperform every frontier model at its own game.
The Problem: No Single Model Does Everything Well
Frontier models each dominate in different areas — one leads in coding, another in scientific reasoning, another in mathematics. Relying on a single model means inheriting all of its weaknesses alongside its strengths.
Sakana AI's answer is Fugu: a multi-agent system that routes each task to the model best equipped to handle it, then compiles a final result stronger than any single model could produce alone.
What is Sakana AI?
Sakana AI is a Tokyo-based company founded in July 2023 by former Google AI researchers. "Sakana" is Japanese for fish — and "Fugu" means puffer fish, a delicacy famous for its flavor but also its toxicity. At its launch event, Sakana released:
Fugu
The base multi-agent model. Balances latency and quality — strong default for everyday coding and interactive work.
Fugu Ultra
Prioritizes answer quality on complex, multi-step reasoning — coordinates more expert agents, slower but more accurate.
Mythos Preview
A separate model teased at the same event — details still emerging.
How Fugu Works
Fugu is a front-end orchestrator. It decides whether to answer a query directly using its own reasoning, or to distribute subtasks to other models — GPT, Claude, Gemini, open-source models, or Sakana's own fine-tuned models — before compiling the final response.
Fugu — The Manager
Routes each task to the most capable model based on what the task requires. No single vendor dependency — works with closed and open-source models alike. Fugu itself can also answer directly when it has the capability.
Fusion — The Judge
A different approach available on OpenRouter: calls multiple models in parallel, judges which output is best, then returns that result. Fugu and Fusion are complementary, not the same system.
When multiple agents are active under the pay-as-you-go plan, you pay a single rate based on the top-tier model involved — not each individually.
Benchmark Performance
According to Sakana's promotional benchmarks, Fugu Ultra outperforms Claude Fable 5 and sits shoulder-to-shoulder with Mythos Preview:
| Benchmark | Fugu Ultra | Claude Fable 5 | Claude Opus | GPT | Gemini |
|---|---|---|---|---|---|
| Agentic Coding (8-bit) | 82.1% | 80.4% | — | — | — |
| Live Code Bench | 93.2% | 89.8% | — | — | — |
| Long Context Reasoning | 73% | — | 67% | 74.3% | 72% |
| Long Context Reasoning (Fugu) | 70% | — | 67% | 74.3% | 72% |
Benchmarks sourced from Sakana AI promotional materials. Independent third-party verification pending.
Pricing Breakdown
Subscription Plans
| Plan | Price / mo | Usage |
|---|---|---|
| Standard | $20 | Lightweight daily use |
| Pro | $100 | 10× Standard |
| Enterprise | $200 | 30× Standard |
Pay-As-You-Go
| Model | Input /M tokens | Output /M tokens |
|---|---|---|
| Fugu | Standard rate of the underlying model used | |
| Fugu Ultra | $5 | $30 |
| Fugu Ultra (>272K ctx) | $10 | $45 |
| Cached Input | $0.50 | — |
Also available on OpenRouter at the same pricing.
Who Is Using Fugu — and How?
Fugu is already being deployed in high-leverage workflows where task complexity justifies multi-model coordination:
Coding and Code Review
Software engineers route complex review tasks through Fugu, letting the system pull specialized models for different parts of the codebase — syntax checking, logic review, security scanning — in a single coordinated pass.
Research Acceleration
Tasks that would take 3–4 days were completed in a few hours — including reading a paper, implementing the training, evaluating results, and analyzing gaps, all handled autonomously.
End-to-End Security Assessment
Fugu drove a full security assessment autonomously for nearly four hours — from initial reconnaissance through vulnerability analysis to final reporting, with minimal human intervention.
How to access Fugu today
Via OpenAI-compatible API, OpenRouter, or directly through console.sakana.ai. Create an API key, add payment, and integrate with OpenClaw, Codex, or any agentic workflow.
Recommended Deep Dives
Read Our Blogs
How One B2B SaaS Founder Used Genuine Influencer Marketing to Reach 10 Paying Customers Organically
No customer database, no audience, and every personal contact already asked. How one founder used niche creators on Amplibee to go from 2 to 10 paying customers — with no ads and no bought engagement.
Why Cross-Posting Tools Like Postiz and Hootsuite Aren't Enough to Grow Fast
Schedulers like Postiz and Hootsuite are great at distributing posts to the audience you already have. Fast growth needs new, real audiences — which is where genuine influencer marketing comes in.